Re: resend: tcp: performance issue with fastopen connections (mss > window)

[Alexey Kodanev <alexey.kodanev@oracle.com>]

Hi Eric,
On 13.01.2017 18:35, Eric Dumazet wrote:
> On Fri, 2017-01-13 at 18:01 +0300, Alexey Kodanev wrote:
>> Hi,
>>
>> Got the issue when running LTP/netstress test on localhost with mss
>> greater than the send window advertised by client (right after 3WHS).
>> Here is the testscenario that can reproduce this:
> Hi Alexey
>
> So this is a combination of Fastopen + small window + large MSS ?

Yeah, this happens only in the beginning, after first ack from client.
Later window gets
lager than mss and it doesn't happen.

>
> I would rather not force burning tons of coal or other fossil fuel,
> by making each tcp_sendmsg() done by billions of linux devices more
> expensive, only to accommodate for some LTP test doing something not
> sensible ;)
>
> Fact that you removed one condition in the BUG_ON() might hide another
> issue later in the path.
>
> I would suggest to clamp MSS to half the initial window, but I guess
> this is impractical since window in SYN/SYNACK are not scaled.
> Care to send a packetdrill test so that we have a clear picture of what
> is going on ?

Is it capable of making two connections in the single test, one after
another?

Thanks,
Alexey